Qwen 3.6-27B hits 5x speedup via DFlash

// 45d agoMODEL RELEASE

Qwen 3.6-27B hits 5x speedup via DFlash

A new DFlash draft model optimized for Qwen 3.6-27B has been released, resolving the significant performance regressions seen when using older Qwen 3.5 speculators. This release enables high-speed speculative decoding for repository-level agentic coding tasks on local hardware.

// ANALYSIS

Speculative decoding is shifting from a luxury to a requirement for local LLM viability, and the DFlash release for Qwen 3.6 is a critical infrastructure update. Architectural changes in Qwen 3.6, specifically Causal Sliding Window Attention (SWA), broke compatibility with older speculators, causing the "PP speed" issues reported by early adopters. The new model from Z-Lab bridges this gap, offering a 3x-5x throughput increase on consumer-grade GPUs like the RTX 4090. While Qwen 3.6 has native Multi-Token Prediction (MTP), the DFlash speculator remains the gold standard for high-performance inference in specialized engines like vLLM and SGLang. This release makes the 27B model's "agentic coding" and "Thinking Preservation" features truly real-time for local developers.

// TAGS

qwendflashspeculative-decodinglocal-llmllm

DISCOVERED

45d ago

2026-04-25

PUBLISHED

45d ago

2026-04-25

RELEVANCE

8/ 10

AUTHOR

butterfly_labs

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

NEWS13m ago

Apple Siri reportedly uses Gemma 4

While Apple's new Siri AI is not based on Google's proprietary Gemini model, it reportedly utilizes a customized version of Gemma 4 - E4B, a smaller open-source model developed by Google. To optimize execution on consumer hardware with limited RAM, Apple employs a specialized per-request Mixture of Experts (MoE) scheme that loads model weights directly from NAND flash memory.

LAUNCH15m ago

HeyGen has released Hyperframes as an official Claude connector, enabling AI agents and users to generate and render videos programmatically using standard web technologies.

HeyGen has introduced Hyperframes, an open-source framework and Model Context Protocol (MCP) server that integrates directly with Claude to compile HTML, CSS, and JavaScript into rendered videos. By partnering with Anthropic, HeyGen addresses the limitation of text-dense LLM outputs by allowing AI agents to generate structured, visual MP4 videos in real-time. This "video as code" approach makes the creation of product walkthroughs, animated explainers, and marketing clips highly programmatic and fully customizable.

TUTORIAL19m ago

OpenRouter has launched a dedicated API override endpoint that allows Cursor users to route requests to hundreds of LLMs through a single API key.

OpenRouter has introduced a dedicated integration endpoint (https://openrouter.ai/api/v1/cursor) to support Cursor, an AI-powered code editor. By overriding Cursor's OpenAI base URL, developers can leverage OpenRouter to access hundreds of AI models with a single API key. This integration normalizes Cursor's unique request format, enabling tool and function calls to run smoothly. It also provides automatic provider failover, centralized cost/budget tracking, and customizable routing configurations for individuals and teams, though tab completions and certain Composer modes remain restricted to Cursor's native setup.

Qwen 3.6-27B hits 5x speedup via DFlash